[hsliuustc0106]

Required human contributor answers

Answer every question before implementation and PR submission, including drafts.
Reuse explicit, still-applicable human answers with a traceable source reference;
ask the human for anything missing and wait. An agent must not invent answers.
Blank fields, placeholders, or an agent-written rationale do not satisfy this gate.
Share only information and source links appropriate for this PR's audience.

1. What outcome are you trying to achieve?

   Answer: Add a two-tier Public231 evaluation path so tips that clearly regress are dropped without a full C1, and record the measured StartLux Instant screen primary. Source: Hongsheng's authorizing task (2026-10-25), the 2026-10-26 WITHHOLD fix, InferRSI's quota decision, and the hk2 n=58 screen measurement the same day.

2. What is the highest-leverage variable for that outcome?

   Answer: A fast-screen after smoke whose baseline is the primary SHA's own screen of the same request IDs on the same machine, never the full-C1 pooled pair. Quotas are 10/10/8/8/6/16 (58 when every bucket is full). A short bucket records quota_shortfall and is not filled from another bucket. 2048+ below 16 is inconclusive and cannot early-DROP. Early DROP only when both of those screen percentiles are worse by ≥ +2% or ≥ +0.5 ms. Source: InferRSI protocol decision, system1 room 2026-10-26.

3. How will this work move that variable rather than create busywork?

   Answer: Document the rule in the existing skills and wire it through benchmarks/bench.py --fast-screen / --baseline-screen, refusing typed pooled percentiles. Record the hk2 smoke PASS and the measured Instant screen primary below. Do not type those screen percentiles in place of --baseline-screen. Source: same decision and the hk2 measurement report.

4. Who is the customer or user, and what do they actually need?

   Answer: Maintainers and the InferRSI loop. They need a clear regression to stop before full C1, and KEEP to stay on the full dual gate. Source: Hongsheng authorized the update; InferRSI set the quotas; hk2 measured the screen primary.

After answering the four questions: is a PR needed at all to achieve this outcome?

Human decision and rationale (or explicit source reference): Yes. Open and keep draft PR #181. Do not mark ready and do not merge.

For a bug fix: link actual observed failure evidence.
Include the observed result of a minimal reproduction, failing regression test,
or relevant trace. Label untested diagnoses as unverified. For a non-bugfix,
state why this is not applicable.

Evidence: Not a bug fix. The numbers below are the operator-reported hk2 measurement. This commit does not remeasure them.

Purpose

Screen baseline is the same primary SHA, same machine, same request IDs. It is never a full-C1 pooled p50/p95. --baseline-p50-ms and --baseline-p95-ms are refused. A full-C1 directory is refused. StartLux Instant pooled 7.338/51.816 (51f80ad681c8f7b1be9b22b5e700bef34e33534c) is refused as a screen baseline.

Quotas are 0–127:10, 128–255:10, 256–511:8, 512–1023:8, 1024–2047:6 (was 8), 2048+:16. That is 58 requests when every bucket is full. Seed 231, one round, concurrency 1, warmup 5.

Cap-to-available: if a bucket has fewer prompts than its quota, take every prompt in that bucket and record quota_shortfall. Do not fill from another bucket. Do not substitute another model's lengths. If 2048+ has fewer than 16, the whole screen is inconclusive and cannot early-DROP.

DROP only when both screen deltas are worse (tip >= baseline * 1.02 or tip >= baseline + 0.5 ms) against that same-ID screen primary. A DROP writes dropped_at (YYYY-MM-DDTHH:MM:SSZ) in fast_screen_drop.json and on the summary. KEEP is still only the full C1 dual gate.

StartLux Instant screen primary (n=58, quotas 10/10/8/8/6/16), reported from hk2 and not remeasured here:

screen_primary_sha: 51f80ad681c8f7b1be9b22b5e700bef34e33534c
screen_primary_p50_ms: 19.112976267933846
screen_primary_p95_ms: 73.46072793006897
n: 58
successful_requests: 58/58
evidence_complete: true
fast_screen_decision: screen_primary
quota_shortfall: {}
bucket_counts: 0-127:10, 128-255:10, 256-511:8, 512-1023:8, 1024-2047:6, 2048+:16
device: L20X
compute capability: sm_89

The measuring client was 923c8c739bcaf3cc2b098f3cf373c2b547d2666c plus a local-only patch that changed the 1024–2047 quota from 8 to 6. The remote tip may still have been on quota 8 when this was measured. This branch's committed runner samples n=58 with quota 6, so the next screen does not need that local patch. Artifacts on hk2: ~/experiments/2026-10-26-pr181-screen-primary/screen-primary-n58/ (SCREEN_PRIMARY.json, fast-screen/, screen_ids.json, local-quota-patch.diff).

Do not type 19.112976267933846 / 73.46072793006897. Pass the saved run as --baseline-screen. Do not write 7.338/51.816 as this screen primary.

hk2 P2 contract smoke PASS, separate from the screen: client d2fab1e08f231fe2e23540f3b25b2650cd416d09 against StartLux Instant served at 51f80ad681c8f7b1be9b22b5e700bef34e33534c, 4/4, evidence_complete true. Artifacts: ~/experiments/2026-10-26-pr181-screen-primary/primary/smoke/ on hk2. That client has no StartLux or CUDA diff versus 51f80ad. Offline tests/benchmarks on that client: 109 OK. The 4/4 smoke is not the screen primary and not a latency pair.

Clef and pplx-decider keep their full-C1 primaries in the recipes and the latency leaderboard. They do not inherit Instant's screen pair.

Test Plan

System1-Omni Version / Commit: f675976db36167d0ec6911805990c184d21701c9 on cursor/public231-fast-screen-3f8c, base a41e379453d4a8dc3cbe3f76c8754753c42230a1.

• python3 -m unittest discover -s tests/benchmarks -p 'test_*.py' -q

No new GPU run in this commit. The hk2 4/4 smoke and the n=58 screen primary are operator-reported.

Test Result

• Local python3 -m unittest discover -s tests/benchmarks -p 'test_*.py' -q: 112 tests, OK. Typed 7.338/51.816 is still refused. A short bucket records quota_shortfall. 2048+ below 16 is inconclusive and writes no dropped_at.
• Operator-reported on hk2: smoke 4/4 for d2fab1e versus Instant at 51f80ad; screen primary n=58, p50 19.112976267933846 ms, p95 73.46072793006897 ms, evidence_complete true, decision screen_primary, quota_shortfall {}.
• Not produced here: a second measurement of those screen percentiles.

Demo / evidence

N/A as a new latency campaign. The screen primary and the smoke PASS are the operator-reported hk2 artifacts above. This PR stays draft.

Self-review

Before marking this PR ready for review or requesting maintainer review, complete
the self-review checklist.
Keep the PR in draft while this work is incomplete.
For agent assistance, use the optional precheck-pr skill.

• I have reviewed the full diff and addressed the issues I found.
• I have checked that the change follows the project's architecture and stays focused on the stated purpose.
• I have run the checks appropriate to this change and reported commands, results, and anything I could not verify above.
• I have checked that the PR description, documentation, and any accuracy or performance claims match the implementation and available evidence.

Precheck notes: head f675976db36167d0ec6911805990c184d21701c9. Screen primary is the reported n=58 pair. Do not mark ready.